How do I use IORING_OP_READ_FIXED?

Question

On the following page https://lwn.net/Articles/810414/

IORING_OP_READ_FIXED IORING_OP_WRITE_FIXED These opcodes also submit I/O operations, but they use "registered" buffers that are already mapped into the kernel, reducing the amount of total overhead.

However I could not find a single example online on how to use it. In io_uring_enter it says

EFAULT IORING_OP_READ_FIXED or IORING_OP_WRITE_FIXED was specified in the opcode field of the submission queue entry, but either buffers were not registered for this io_uring instance, or the address range described by addr and len does not fit within the buffer registered at buf_index.

It seems to me that I should pick a memory address and block for it to use but using an address like 0x555555500000 and len as 4096 gets me the same error.

How does IORING_OP_READ_FIXED work? Below is a working example of IORING_OP_READ

#include <stdio.h>
#include <stdlib.h>
#include <fcntl.h>
#include <unistd.h>
#include <string.h>
#include <sys/stat.h>
#include <sys/syscall.h>
#include <sys/mman.h>
#include <linux/io_uring.h>

#define read_barrier()  __asm__ __volatile__("":::"memory")
#define write_barrier() __asm__ __volatile__("":::"memory")

int main(int argc, char *argv[])
{
    struct io_uring_params uring;
    memset(&uring, 0, sizeof(uring));
    auto queue_size = 5;
    auto ring_fd = syscall(__NR_io_uring_setup, queue_size, &uring);

    auto*uring_ptr = (char*)mmap(0, uring.sq_off.array + uring.sq_entries * 4, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, ring_fd, IORING_OFF_SQ_RING);
    auto*submit_entries = (io_uring_sqe*)mmap(0, uring.sq_entries * sizeof(struct io_uring_sqe), PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, ring_fd, IORING_OFF_SQES);

    unsigned &sqHead = *(unsigned*)(uring_ptr + uring.sq_off.head);
    unsigned &sqTail = *(unsigned*)(uring_ptr + uring.sq_off.tail);
    unsigned &sqMask = *(unsigned*)(uring_ptr + uring.sq_off.ring_mask);
    unsigned &sqFlags = *(unsigned*)(uring_ptr + uring.sq_off.flags);
    unsigned *sqArray = (unsigned*)(uring_ptr + uring.sq_off.array);

    unsigned &cqHead = *(unsigned*)(uring_ptr + uring.cq_off.head);
    unsigned &cqTail = *(unsigned*)(uring_ptr + uring.cq_off.tail);
    unsigned &cqMask = *(unsigned*)(uring_ptr + uring.cq_off.ring_mask);
    io_uring_cqe *cqes = (io_uring_cqe*)(uring_ptr + uring.cq_off.cqes);


    int fd[2];
    fd[0] = open(argv[1], O_RDONLY);
    

    struct stat stat;
    if (fstat(fd[0], &stat) < 0) {
        perror("fstat");
        return -1;
    }

    int size_aligned = (stat.st_size & ~63) + (stat.st_size & 63 ? 64 : 0);

    auto*fileBuf = (unsigned char*)malloc(size_aligned*2);

    for(int i=0; i<1; i++)
    {
        io_uring_sqe&sqe = submit_entries[sqTail & sqMask];
        sqe.fd = fd[i];
        sqe.flags = 0;
        sqe.opcode = IORING_OP_READ;
        sqe.addr = (unsigned long long)fileBuf+i*size_aligned;
        sqe.len = size_aligned;
        sqe.user_data = (unsigned long long)fileBuf+i*size_aligned;
        sqArray[sqTail&sqMask] = sqTail&sqMask;
        sqTail++;
    }
    write_barrier();

    //int ret =  syscall(__NR_io_uring_enter, ring_fd, 2, 2, IORING_ENTER_GETEVENTS, 0);
    int ret =  syscall(__NR_io_uring_enter, ring_fd, 1, 1, IORING_ENTER_GETEVENTS, 0);
    //int ret =  syscall(__NR_io_uring_enter, ring_fd, 1, 0, IORING_ENTER_GETEVENTS, 0);
    //sleep(1);
    read_barrier();
    while (cqHead != cqTail)
    {
        unsigned long long a = cqHead;
        unsigned long long b = cqTail;
        unsigned long long c = cqMask;
        auto index=cqHead & cqMask;
        io_uring_cqe&cqe = cqes[index];
        auto u=cqe.user_data;
        auto f=cqe.flags;
        auto r=cqe.res;
        puts((const char*)u);
        cqHead++;
    }
    int a=0;
    return 0;
}

Did you register the buffer for the io_uring instance? When I search Google for 'io_uring register buffer' I get some information about registering buffers. — user253751, Jan 04 '21 at 10:18

score 3 · Accepted Answer · answered Feb 02 '21 at 00:42

You need to register (and unregister) your buffers with the __NR_io_uring_register system call. Using your example, you would apply your buffer to an iovec and the pass the iovec into the syscall:

struct iovec iov = { .iov_base = (void *)fileBuf, .iov_len = (size_aligned*2) };

int rc = syscall(__NR_io_uring_register, ring_fd, IORING_REGISTER_BUFFERS, 
                 (void *)&iov, 1 /* number of iovs */);

To use the registered buffer you need to use provide the buffer's array offset in the sqe. In this example, that value would always be '0' since only one iovec has been registered. In your code, you would need to set sqe.opcode = IORING_OP_READ_FIXED and sqe.buf_index = 0.

You may also want to consider using the initialization helper functions from liburing.h:

static inline void io_uring_prep_read_fixed(struct io_uring_sqe *sqe, int fd,
                                            void *buf, unsigned nbytes,
                                            off_t offset, int buf_index)
{
        io_uring_prep_rw(IORING_OP_READ_FIXED, sqe, fd, buf, nbytes, offset);
        sqe->buf_index = buf_index;
}

I recommend looking at liburing - it aptly handles much of the tedious initialization minutiae as well providing wrappers for buffer [un]registration.

How do I use IORING_OP_READ_FIXED?

1 Answers1